Garbage collection method is not manageable by environment variables in .net 6

[erexo]

Describe the bug

Documentation mentions that user is able to change the default Server garbage collection to Workstation garbage collection via either COMPlus_gcServer (pre .net6) or DOTNET_gcServer (in .net6) Environment variable.
It was working (the COMPlus env) in .net5, but after project upgrade to .net6, both methods seems to be doing nothing, gcServer is always set to true.
MSBuild property is still functioning, we can add /p:ServerGarbageCollection=false to the build parameters and we will achieve exactly that.
The problem seems to be asp.net related, Console application (with default gc set to Workstation) can be overridden by both of those env variables to Server.

To Reproduce

• Create webapi template dotnet new webapi in .net6
• Run that project with any of given env variables (COMPlus_gcServer=0 or DOTNET_gcServer=0)
• Check used garbage collection method (ie. print System.Runtime.GCSettings.IsServerGC on startup)

Further technical details

• ASP.NET Core version: 6.0
• The IDE: VS Code 1.62.2-1636665017_amd64

dotnet --info Output

NET SDK (reflecting any global.json):
 Version:   6.0.100
 Commit:    9e8b04bbff

Runtime Environment:
 OS Name:     ubuntu
 OS Version:  20.04
 OS Platform: Linux
 RID:         ubuntu.20.04-x64
 Base Path:   /usr/share/dotnet/sdk/6.0.100/

Host (useful for support):
  Version: 6.0.0
  Commit:  4822e3c3aa

[dotnet-issue-labeler]

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

[ghost]

Tagging subscribers to this area: @dotnet/gc
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Describe the bug

Documentation mentions that user is able to change the default Server garbage collection to Workstation garbage collection via either COMPlus_gcServer (pre .net6) or DOTNET_gcServer (in .net6) Environment variable.
It was working (the COMPlus env) in .net5, but after project upgrade to .net6, both methods seems to be doing nothing, gcServer is always set to true.
MSBuild property is still functioning, we can add /p:ServerGarbageCollection=false to the build parameters and we will achieve exactly that.
The problem seems to be asp.net related, Console application (with default gc set to Workstation) can be overridden by both of those env variables to Server.

To Reproduce

• Create webapi template dotnet new webapi in .net6
• Run that project with any of given env variables (COMPlus_gcServer=0 or DOTNET_gcServer=0)
• Check used garbage collection method (ie. print System.Runtime.GCSettings.IsServerGC on startup)

Further technical details

• ASP.NET Core version: 6.0
• The IDE: VS Code 1.62.2-1636665017_amd64

dotnet --info Output

NET SDK (reflecting any global.json):
 Version:   6.0.100
 Commit:    9e8b04bbff

Runtime Environment:
 OS Name:     ubuntu
 OS Version:  20.04
 OS Platform: Linux
 RID:         ubuntu.20.04-x64
 Base Path:   /usr/share/dotnet/sdk/6.0.100/

Host (useful for support):
  Version: 6.0.0
  Commit:  4822e3c3aa

Author:	Erexo
Assignees:	-
Labels:	area-GC-coreclr, untriaged
Milestone:	-

[davidfowl]

I can reproduce this one. I'm wondering what happens if you have both the environment variable and the value set in runtimeconfig.json

[mangod9]

so are you observing that the value in runtimeconfig.json is now overriding what is set in the env. variable?

[cshung]

I can reproduce this one. I'm wondering what happens if you have both the environment variable and the value set in runtimeconfig.json

In general, the design is that the environment variable should override the settings in runtimeconfig.json.

The GCToEEInterface::GetBooleanConfigValue function do just that.

Unfortunately, the gcServer is a special case because we need to read that configuration before we initialize the heap, so the reading of the flag is outside of this function. The g_heap_type is set in InitializeStartupFlags in ceemain.cpp.

The startupFlags are set in SetStartupFlags which is setting a value computed earlier in InitializeStartupFlags in unixinterface.cpp, that function uses Configuration::GetKnobBooleanValue which also make sure we read environment variable first.

Not sure what is going wrong - @davidfowl, can you help us to debug what is happening in these two functions? Maybe we can find the bug there.

[mangod9]

The startupFlags are set in SetStartupFlags which is setting a value computed earlier in InitializeStartupFlags in unixinterface.cpp

So does this only repro on linux or windows too?

[davidfowl]

Windows as well.

[cshung]

I have found the culprit of this bug, it appears to be this line:

runtime/src/coreclr/utilcode/clrconfig.cpp

Line 453 in cc44d43

// Ignore the default value even if it's set explicitly.

So first thing first, while confusing, unixinterface.cpp is actually involved in Windows as well. We really should rename it.

Following the analysis above, when the runtime startup, it uses Configuration::GetKnobBooleanValue to determine whether or not to use server GC, it is obvious that the intent of that function is that the environment variable can override the runtimeconfig.json

runtime/src/coreclr/utilcode/configuration.cpp

Lines 108 to 121 in cc44d43

	bool Configuration::GetKnobBooleanValue(LPCWSTR name, const CLRConfig::ConfigDWORDInfo& dwordInfo)
	{
	bool returnedDefaultValue;
	DWORD legacyValue = CLRConfig::GetConfigValue(dwordInfo, &returnedDefaultValue);
	if (!returnedDefaultValue)
	{
	return (legacyValue != 0);
	}
	LPCWSTR knobValue = GetConfigurationValue(name);
	if (knobValue != nullptr)
	{
	return (wcscmp(knobValue, W("true")) == 0);
	}

In the majority of the cases, the code is fine, because there is really no point for someone to specify an environment variable with value that is exactly the default value (for the runtime).

Unfortunately, if the user wanted to disable server GC in WebAPI, then the user would set the value to be 0, which also happened to be the default value for the runtime. Now the culprit line of CLRConfig::GetConfigValue would determine while the environment variable is already set, but it is setting to a default value, so ignoring the fact that it is set and set *isDefault to true anyway. As such, Configuration::GetKnobBooleanValue would move on and read the runtimeconfig.json anyway, even when the environment variable is set.

Apparently, that behavior is not what we want in this issue.

Given the analysis, it isn't difficult to fix this one, but I wonder why one put in that specific ignore there. In particular, I don't want to regress something else. @AaronRobinsonMSFT, do you know why we want to specifically ignore the cases where the user explicitly specified the environment variable with exactly the default value?

[AaronRobinsonMSFT]

@AaronRobinsonMSFT, do you know why we want to specifically ignore the cases where the user explicitly specified the environment variable with exactly the default value?

If I recall that logic was previously there from long back https://github.com/dotnet/coreclr/blob/1d483901530c2d5b8a943fab932e8c184d2c5f36/src/utilcode/clrconfig.cpp#L295-L300

I tried to maintain the identical behavior but it seems there was a branch in the GC case that I missed. The old logic was so intertwined with registry access I'm not sure which part has changed though.

[AaronRobinsonMSFT]

I would fix this outside of the clrconfig.cpp file. One can use the following to query if it is set.

runtime/src/coreclr/utilcode/clrconfig.cpp

Line 580 in cc44d43

BOOL CLRConfig::IsConfigOptionSpecified(LPCWSTR name)

[mangod9]

Thanks for investigating @cshung. We should make a fix in 7 and will possibly need to port to 6. Is the workaround to just remove the setting from runtimeconfig.json then?